Huang et al. (2023) — ARXIV-2310.01798
Key takeaways
- The distinction is the contribution. Intrinsic self-correction is a model revising its answer on its own inherent capabilities alone, with no external feedback, no oracle, no verifier, no second model. Results claimed for “self-correction” often quietly include one of those, and the two situations do not behave alike.
- Under that definition, on reasoning tasks, models struggle to improve and at times degrade after being asked to revise.
- The failure is not a prompting detail. A model with no external signal has no way to tell a wrong answer from a right one, so a revision step is as likely to move a correct answer as an incorrect one.
Standing in the anthology
Filed as the negative result a later practice is built on top of, which is the honest way to hold it. LIT-383 (SCoRe) exists because this one is true: it does not dispute the finding, it changes the intervention from prompting to training, and reports its gains against exactly the baseline this paper establishes.
Carries no practice of its own. What it licenses is a condition on somebody else’s — the reason SOTA-233 says to train the capability rather than ask for it — and a warning about reading any self-correction claim without first checking whether external feedback was in the loop.